Tested in environment:
- Conda
- PyTorch 1.4
- CUDA 10.1
- Ubuntu 18.04
- Python 3.7
- Creating pytorch environment:
conda create -n pvrcnn python=3.7
conda activate pvrcnn
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
- Installing Pointnet2:
git clone https://github.com/sshaoshuai/Pointnet2.PyTorch.git
cd Pointnet2.PyTorch && python setup.py install
export PYTHONPATH=$PYTHONPATH:/path/to/Pointnet2.PyTorch/
- Installing patched spconv:
git clone https://github.com/jhultman/spconv.git --recursive
cd spconv && python setup.py bdist_wheel
cd ./dist && pip install *.whl
- Installing torchsearchsorted:
git clone https://github.com/aliutkus/torchsearchsorted.git
cd torchsearchsorted && pip install .
- Installing misc python packages:
pip install -r requirements.txt
- Installing pvrcnn (this package):
python setup.py develop
If the custom ops are not compiled with GPU support, make sure your
CUDA_HOME
environment variable is set, and try running setup.py
with FORCE_CUDA=1
.